WeDLM Diffusion Language Model Tutorial: 3× Faster Inference Than vLLM AR Models

Background

In large‑scale deployment and commercial scenarios, inference speed often outweighs raw model size as the key factor for engineering value. Autoregressive (AR) generation produces tokens one‑by‑one, which prevents effective use of parallel compute and leads to high latency and cost, especially for long texts, complex reasoning, and high‑concurrency services.

Motivation for Diffusion Language Models

Recent research has explored parallel decoding paths, and diffusion language models (DLMs) are attractive because they can generate multiple tokens per step. However, many existing DLMs rely on bidirectional attention, which destroys the prefix key‑value (KV) cache that modern inference engines such as vLLM depend on. The broken cache forces repeated recomputation of the context, cancelling the potential speed gains of parallel generation.

WeDLM Design

The Tencent WeChat AI team proposes WeDLM (WeChat Diffusion Language Model), the first diffusion model that outperforms an AR model when run on the industrial‑grade vLLM engine. WeDLM keeps a strict causal mask while allowing every masked position to condition on all previously observed tokens. The core technique is Topological Reordering , which moves observed tokens into the physical prefix region without changing their logical order, thereby preserving the KV cache.

Experimental Results

Experiments show that WeDLM maintains the generation quality of strong AR backbones and achieves substantial inference acceleration. On mathematical reasoning tasks, WeDLM is more than three times faster than an AR model deployed with vLLM. In low‑entropy scenarios, the speedup exceeds ten times.

Resources

Source code and tutorial are available at https://github.com/tencent/WeDLM. The framework can be run on a GPU‑accelerated environment (e.g., NVIDIA GeForce RTX 5090) using the vLLM inference engine.